Four main steps of our methodology. (1) Deep learning model training using 24 hr temporal sequences of SHARP parameters and GOES ≥M-class flare occurrence as targets. (2) Feature extraction from input sequences, including parameter values at each time step, their temporal derivatives, and associated statistics. (3) Feature selection: top five SHARP parameters from permutation-based evaluation on the best deep learning model; top 15 features by impurity-based feature importance (mean decrease in impurity) from a random forest trained on features derived from those parameters. (4) SR to estimate interpretable mathematical mappings from selected features to calibrated deep learning probability estimates.